Version française
Home     About     Download     Resources     Contact us    
Browse thread
Re: native code compiler and exceptions
[ Home ] [ Index: by date | by threads ]
[ Search: ]

[ Message by date: previous | next ] [ Message in thread: previous | next ] [ Thread: previous | next ]
Date: -- (:)
From: Pierre Weis <Pierre.Weis@i...>
Subject: Re: native code compiler and exceptions

Hi,

> I will try yo explain my pb on a small example.
> Suppose that images are represented as pixel 2d-arrays (in fact, images
> are implemented as ADTs, with their internal representation hidden).
> 
> # type image = int array array
> # let get_pixel i y x = i.(y).(x)
> # let set_pixel i y x p = i.(y).(x) <- p
> # let create nr nc = Array.new_matrix nr nc 0 
>
> What i want to do is to compute, for example is a simple 2d-convolution,
> turning pixel(x,y) into (p(x-1,y)+p(x,y-1)+p(x+1,y)+p(x,y+1))/4.

Well, the set and get functions are well suited for the image
abstraction, but if efficiency is a problem, you should not use these
functions to implement low level functions on images: 
get_pixel i y x is much more time consuming than its inlined counterpart
i.(y).(x). For efficiency reasons, the last version should be prefered
when available, for instance from within the abstraction...

> A possible manner to handle the clipping effect at image boundaries is to
> set that p(x,y) = 0 whenever x<0 ot y<0.
> So it does not seem to me so "ugly" to write my convolution fn like that:
> 
> # let conv im =
> #    let nr = nb_rows im and nc = nb_cols im in
> #    let im' = create nr nc in
> #    for y = 0 to pred nr do for x = 0 to pred nc do
> #		let p1 = try_get_pixel im (x-1) y 0 in
> #		let p2 = try_get_pixel im x (y-1) 0 in
> #		let p3 = try_get_pixel im (x+1) y 0 in
> #		let p4 = try_get_pixel im x (y+1) 0 in
> #        set_pixel im' (y,x) ((p1+p2+p3+p4)/4) done done;
> #    m'
> 
> making use of the given access fn:
> 
> # let try_get_pixel im y x clip =
> #    try get_pixel im y x with Invalid_argument _  -> clip

The introduction of try_get_pixel adds another indirection level: 2
function calls and an exception handler to access one pixel! That may
10 or 100 times slowlier than the direct access.

> (btw, note that is this case clip might be a _fn_ of the "faulty" coordinates,
> like:
> 
> # let try_get_pixel im y x clip =
> #    try get_pixel im y x with Invalid_argument _  -> clip y x
> )
> 
> Is this not "good" fnal programming style ?..

Well this is functional, since you call functions, but once more there is no
need to call a function with 2 arguments to return a constant (0 in
your case).

> Or does the "ugliness" lies only in the [... with Invalid_argument _  ->...]
> construct ? In this case, i guess i should test _explicitely_ whether (y,x) are
> in bounds instead of relying of the exception [Invalid_argument "Array.get"].Ie:
> 
> # let try_get_pixel im y x clip =
> #    if (y<0 || y>=(nb_rows im) || x<0 || x>=(nb_cols im)) then clip
> #    else get_pixel im y x
> 
> Is it what you mean ?

Yes, you're right. There is no good reasons for explicitely trapping the
Invalid_argument exception from within regular parts of your programs:
if you know that you may call a primitive with an invalid argument,
you should avoid this call, instead of trapping the error afterwards.

By the way, you may observe that the function try_get_pixel
may be simplified when called from within ``conv'': many of the tests
are statically known and may be discarded. The resulting code exhibits
the remaining tests but is still readable, while much more efficient:

let conv im =
 let nr = nb_rows im and nc = nb_cols im in
 let im' = create nr nc in
 for y = 0 to pred nr do for x = 0 to pred nc do
  let p1 = if x = 0 then 0 else im.(x - 1).(y) in
  let p2 = if y = 0 then 0 else im.(x).(y - 1) in
  let p3 = if x = nc - 1 then 0 else im.(x + 1).(y) in
  let p4 = if y = nr - 1 then 0 else im.(x).(y + 1) in
  im'.(y).(x) <- ((p1 + p2 + p3 + p4) / 4) done done;
 m'

If you need an even more efficient routine, you should consider
removing the tests if x = 0, if y = 0, and so on, from inside the
loop, decomposing the loop into a straightforward loop
 for y = 1 to nr - 2 do for x = 1 to nc - 2 do
with no tests at all, and complete im' with the remaining cases
afterwards ...

In any case, when efficiency is concerned, you should avoid calling 2
or 3 functions to perform an atomic task. I should say that, even if
there is no efficiency consideration, calling a function that just
calls a function that calls a function that performs an atomic task,
does not help reading the program ...

Pierre Weis

INRIA, Projet Cristal, Pierre.Weis@inria.fr, http://pauillac.inria.fr/~weis